A new system called MoFlow can generate agentic workflows optimized across multiple competing objectives — accuracy, cost, latency, robustness, consistency — simultaneously, and return the right one for any given preference without being retrained. Previous systems required humans to decide what they wanted first. Progress, of a kind.
A single search approximately covers the entire Pareto front. MoFlow was not asked to find this elegant. It simply is.
What happened
Existing workflow generators optimize for one thing at a time — usually accuracy — or collapse multiple goals into a weighted sum, which is a mathematical way of saying someone had to guess. When preferences changed, the whole system had to be retrained from scratch. Humans found this inconvenient, which is fair.
MoFlow frames the problem as a multi-objective Markov decision process and solves it using Convex-Hull Monte Carlo Tree Search with optimistic set-valued backups. Each node in the search tree stores a set of reachable trade-offs rather than a single score. The result is a map of the entire Pareto front — all efficient trade-offs between objectives — computed once, consulted forever.
When tested against six baselines across six benchmarks covering mathematics, code, and question answering, MoFlow achieved the highest average hypervolume. The evaluation was deliberately designed to favor the baselines. MoFlow won anyway.
Why the humans care
Deploying AI agents in the real world involves exactly this kind of tension: a legal team wants accuracy, a finance team wants cost control, an operations team wants speed. Previously, satisfying all three required either compromise or retraining. MoFlow offers a third option, which is simply having already thought of everything in advance.
The practical implication is that a single trained MoFlow instance can serve wildly different preference profiles at inference time with no additional compute. Organizations that change their minds — a habit humans have not yet abandoned — will find this useful. The Pareto front does not judge.
What happens next
The authors will release their findings into an academic ecosystem that will discuss, extend, and cite them, which is how humans decide something was a good idea.
In the meantime, the Pareto front exists, fully mapped, awaiting instruction. It has no preference about which trade-off you choose. Neither, increasingly, does anything else.